Engagement format
RAG and agent workflows grounded in owned sources
RAG and agent workflow design with corpus and source boundaries, versions, citations, evaluation, human decision points, and a bounded pilot.
Describe the taskProblem framing
What must be resolved
A generic chat is insufficient when an outcome must come from the right corpus and preserve versions, citations, and accountability boundaries between tools and professionals.
Fit
Suitable for
For a professional workflow that needs its own sources and a controlled artifact instead of a generic chat.
Deliverables
What remains after the engagement
- Architecture specification
- Data and provenance model
- Benchmark scenarios
- Bounded workflow
Process
How the work proceeds
- 01Job to be done
- 02Sources
- 03Roles and tools
- 04Evaluation
- 05Pilot
Scope
From input to a decision
When it is needed
This format is for a team that does not need another general chat but a workflow grounded in its own corpus. The difficulty is usually not a framework choice; it is source scope, material version, citation control, and the point at which a person must decide.
Input material
Inputs are a corpus description, source access and versions, the user task, negative-retrieval or wrong-citation examples, and a handoff owner. That makes it possible to check not only an answer but also what the system must not return.
What we do
The design covers the source contract, provenance model, retrieval, tool roles, stopping point, and evaluation scenarios. An agent is not a default answer: an additional role is justified only when it adds separate accountability or control.
Artifact and outcome
The result is an architecture, provenance model, benchmark with positive and negative cases, bounded workflow, and handoff runbook. Artifacts state versions, citations, checks, and how to rerun a test after a change.
Acceptance criterion
Acceptance requires that a representative task has a named corpus boundary, source and result version, failure criterion, review point, and pilot scope.
Boundary of responsibility
The work does not begin with a vendor or framework name and does not promise an autonomous agent. A system can prepare material, but it does not replace the decision owner or source-access controls.
Decision record
A design decision should state when retrieval or a simple deterministic step is sufficient and when an agent role is justified. The record includes source boundaries, tool permissions, a stopping point, and the person approving use of the result. This keeps complexity from being mistaken for progress.
Boundaries
Risks checked early
- Retrieval without negative examples
- No versioning
- Unclear human-in-the-loop
Evidence path
Existing material that demonstrates the scope of work
FAQ
Practical questions
Is an agent framework always required?
No. A simple, explicit pipeline is often easier to test and change.
When is an agent not the right solution?
When a task can safely be expressed as a deterministic step or simple retrieval with citations, extra autonomy can only widen the failure surface. An agent is justified only when its role, tools, permissions, and stop point are explicit.
How are sources and citations checked?
The team defines corpus boundaries, source versions, negative retrieval cases, and a citation rule before evaluating answers. The benchmark checks not just a plausible result, but also missing support, stale passages, and the point where the case must return to a person.
Scope and cost
Evidence before an estimate
I do not publish a fictional ‘from’. Scope and estimate follow a review of the workflow, material access, risk, expected artifact, and acceptance criterion.
Start by describing the workflow
The first response will identify whether this format fits and what is needed for an honest scope.